RFR: 8011218 Kitchensink hanged, likely NMT is to blame

[Daniel D. Daugherty]

Thanks.

Dan


On 4/24/13 4:14 PM, Zhengyu Gu wrote:
> Thanks, Dan.
>
>>    What motivated the new: #include "runtime/thread.inline.hpp"
> This is for Thread::current()
>
> -Zhengyu
>
>
> On Apr 24, 2013, at 6:06 PM, "Daniel D. Daugherty" <daniel.daugherty at oracle.com> wrote:
>
>> On 4/23/13 8:23 AM, Zhengyu Gu wrote:
>>> Round 2:
>>>
>>> Based on feedback from David, Dan and Karen, transition thread state to _thread_in_native is dangerous. This fix keeps query  thread state in _thread_in_vm, but make locks safepoint aware and explicitly check for safepoint during lengthy baseline operation.
>>>
>>> Webrev: http://cr.openjdk.java.net/~zgu/8011218/webrev.01/
>> Thumbs up.
>>
>> src/share/vm/services/memBaseline.cpp
>>     What motivated the new: #include "runtime/thread.inline.hpp"
>>
>> src/share/vm/services/memBaseline.hpp
>>     No comments.
>>
>> src/share/vm/services/memTracker.cpp
>>     No comments.
>>
>> Dan
>>
>>
>>>
>>> Test:
>>>    - Manual test, ran kitchensink stress tests with additional NMT queries to stress query threads, monitored gc.log.
>>>     After 17 hours, found no long pause between gc messages.
>>>
>>> Thanks,
>>>
>>> -Zhengyu
>>>
>>>
>>>
>>> On Apr 18, 2013, at 9:46 AM, Zhengyu Gu wrote:
>>>
>>>> Hi David,
>>>>
>>>> On 4/17/2013 10:59 PM, David Holmes wrote:
>>>>> On 18/04/2013 7:33 AM, Zhengyu Gu wrote:
>>>>>> The thread that runs NMT query can be blocked on NMT snapshot lock,
>>>>>> which prevents JVM to reach safepoints.
>>>>> How? I don't understand this part.
>>>> Currently, NMT acquires locks without safepoint checks (I assumed that query is coming from _thread_in_native, obviously a mistake). Since NMT query can be lengthy, I still want it to run in _thread_in_native state.
>>>>
>>>>>> NMT query does not change Java heap, it does not have to stop for
>>>>>> safepoints. So the solution is to transition to _vm_in_native state,
>>>>>> before executing NMT query.
>>>>> To add to Dan's comments you don't need the assertion:
>>>>>
>>>>>   81   assert(((JavaThread*)thr)->thread_state() == _thread_in_vm,
>>>>>   82     "Just check");
>>>>>
>>>>> because this will be checked by ThreadToNativeFromVM.
>>>>>
>>>>> However I am far from convinced that it is valid to make this state change. You need to be sure that nothing you call from this method - including debug/tracing code that might only be enabled under certain flags - will assume/expect that you are _thread_in_vm.
>>>> NMT query code acquires two locks, both with no_safepoint_check. Besides assertion and tty->print(), I don't recall it uses any vm specific methods, but is there a way to ensure that?
>>>>
>>>> Thanks,
>>>>
>>>> -Zhengyu
>>>>
>>>>
>>>>> Thanks,
>>>>> David
>>>>> -----
>>>>>
>>>>>
>>>>>> JBS: https://jbs.oracle.com/bugs/browse/JDK-8011218
>>>>>> Public bug: not available
>>>>>> Webrev: http://cr.openjdk.java.net/~zgu/8011218/webrev.00/
>>>>>> <http://cr.openjdk.java.net/%7Ezgu/8011218/webrev.00/>
>>>>>>
>>>>>>
>>>>>> Test:
>>>>>>    - Run Kitchensink stress test on low power machine Linux x86 machine
>>>>>> with GC trace on, stress out NMT query thread by submitting NMT query in
>>>>>> a loop, NMT query executions become very slow, but GC trace continues
>>>>>> generating message, and does not show long pause.
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> -Zhengyu